Toward Specialized Learning-based Approaches for Visual Odometry: A Comprehensive Survey

The efficiency of Visual Odometry (VO) is constructed from various sequential components including feature extraction, feature matching, robust displacement estimation, and so on. Recently, many studies have favored learning-based solutions as alternatives to these components. Concurrently, these solutions offer flexibility to researchers in meeting specific demands for modular enhancements. To this end, this survey paper explores the advancements of learning-based methods and how they can get involved in the traditional VO pipeline. This approach enables step-by-step advancement and deeper exploration of individual VO components via the deep learning lens as well as additional algorithms emerging when applied modularly into a baseline. Moreover, our survey extends into end-to-end methods, which streamline the VO process by directly learning camera motion from images. This holistic approach simplifies the VO pipeline and capitalizes on the advantages of DNNs to implicitly model complex relationships in the data. Ultimately, we delve into various common optimization functions and generalized methods crucial in boosting end-to-end VO models or pipeline performance. By juxtaposing these two approaches, this paper aims to provide a comprehensive overview of the DVO approaches for the main baseline as well as the supporting tasks.